Multilevel Models in Social Epidemiology

“Multilevel” structure:

“Multilevel” models:

• Multilevel models developed in 1980s in demography (Entwistle), statistics (Wong/Mason), education (Bryk/Raudenbush).
• Use depends critically on theory.

Our work with individuals has been useful and productive, but this approach alone clearly will not lead to an effective program of health promotion and disease prevention. A new inititive focusing on the environments in which we live much now become a priority for us all.

• Place-based comparisons of health are revealing (Villermé, Farr, Graunt, Snow, DuBois, many others)

• Communities inherently reflect social dynamics.

• Host-Agent-Environment (physical and social).

• “Population perspective”, contra biomedical individualism.

Neighborhood ‘effects’ on violence, mortality

• Focus on mutual adjustment
• Clustering addressed as nuisance

Extended to CVD, low birthweight, other outcomes

• Random effects implementation
• Exploration of multi-level EMM

• Strong theory, field measurements, sophisticated models, potential mechanisms linked to violent crime.

Multilevel analyses showed that a measure of collective efficacy yields a high between-neighborhood reliability and is negatively associated with variations in violence, when individual-level characteristics, measurement error, and prior violence are controlled. Associations of concentrated disadvantage and residential instability with violence are largely mediated by collective efficacy.

• ~4600 families in high poverty randomized to housing vouchers.

• Generated large differences in exposure to high-poverty neighborhoods.

• 5-year follow-up (2003):

  • No impacts on economic self-sufficiency of mothers.
  • Other outcomes mixed, some positive, some negative.

• Many limitations.

What are the problems?

• Social stratification
• Endogeneity
• Extrapolation
• Spillovers

• Evidence for the income inequality/health link was “slowly dissipating”

• Multilevel studies inconsistent in US.

• Weak evidence from Europe and Asia.

• Individual-level controls matter.

Fixed effects: No.

Random effects: Yes!

Zombie hypothesis…

• 86 multilevel papers on neighborhoods
• 80% cross-sectional designs
• Inconsistencies within and across studies.

“it is not clear how much we are learning, or whether such lessons are improving population health…experimental evidence of neighborhood effects is mixed, and observational studies too often report mere correlations, side-stepping critical effect identification issues. Since epidemiologists have long known that disadvantaged environments are not healthy, the utility of studies that do not face the difficult methodological challenges is questionable”

• Now > 250 papers
• Most still using cross-sectional designs with Census data.
• Emphasized the importance of timing of exposures over the lifecourse.
• Pleas for more diversity in study designs.

• Greater focus on credible study designs.

  • Cluster RCTs
  • Quasi-experiments

• Utilizing longitudinal data to focus on changes in exposure

• Weighting methods to deal with observables and post-exposure covariates

• Extensions to mediation

All fit within the scope of multilevel design and analysis

• Defining assumptions for causal effects of contextual exposures

• Time-varying exposures and confounding

• Conditional vs. marginal effects

…the ITT estimate…can successfully measure the effects of the policy initiative, but is not well suited to capturing neighborhood effects.

• Assessed duration of exposure to neighborhood conditions
• Find benefits of shorter exposure to low-poverty

Random assignment of families to different MTO mobility groups…generates large differences in average neighborhood trajectories

Nonexperimental analyses of the type conducted by CM reintroduce all of the selection bias problems that MTO was designed to overcome.

• Time-varying covariates controlled using IPTW, exposure effects estimated using MSMs.

• Can replicate MTO findings.

• Found significant lagged effect of living in concentrated disadvantage compared with advantage at wave 1

• Moving when young increases college attendance and earnings
• Moving as an adolescent has slightly negative impacts.

… suggests that the duration of exposure to better environments during childhood is an important determinant

…incoming refugees were assigned to neighborhoods with varying levels of disadvantage throughout the country

…As a result, this study attempts to address the challenges of selective migration present in existing studies on neighborhood outcomes.

…refugees who were assigned to more disadvantaged neighborhoods were more likely to develop hypertension, hyperlipidemia, diabetes, and MI in subsequent decades.

Effect sizes were small, representing a 2% increase from baseline rates for each condition…

• Recent review of ‘causal analyses’ of neighborhood effects.

• Much more mixed.

• Evidence of selection and confounding.

• Lots of heterogeneity.

• Stronger evidence for children than adults.

• Multilevel analysis of individual heterogeneity and discriminatory accuracy (MAIHDA)
• Individuals nested within social strata
• Partial pooling of intersectoral identities
• Can reveal heterogeneity beyond simple additive effects.

\[y_{ij} = \beta \gamma_{j} + \mu_{0j} + e_{0ij}\]

\[[\mu_{0j}] \sim N(0, \sigma^{2}_{strata})]\]

\[[e_{0ij}] \sim N(0, \sigma^{2}_{e_{0}})\]